InsertNeRF: Instilling Generalizability into NeRF with HyperNet Modules

Authors: Yanqi Bao, Tianyu Ding, Jing Huo, Wenbin Li, Yuxin Li, Yang Gao

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that this method not only achieves superior generalization performance but also provides a flexible pathway for integration with other Ne RF-like systems, even in sparse input settings.
Researcher Affiliation Collaboration 1State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China yq_bao@smail.nju.edu.cn,{huojing,liwenbin,gaoy}@nju.edu.cn, liyuxin16@smail.nju.edu.cn 2Applied Sciences Group, Microsoft Corporation, Redmond, USA tianyuding@microsoft.com
Pseudocode Yes C.2 PSEUDOCODE Algorithm 1: Training for Insert Ne RF-like Systems Algorithm 2: Testing for Insert Ne RF-like Systems
Open Source Code No Code will be available at: https://github.com/bbbbby-99/Insert Ne RF.
Open Datasets Yes During the evaluation phase, we conduct evaluations using three metrics: PSNR, SSIM, and LPIPS, on well-established datasets such as Ne RF Synthetic, LLFF, and DTU. ... Ne RF Synthetic. The dataset consists of 8 synthetic objects ... Local Light Field Fusion (LLFF). The dataset consists of 8 complex real-world scenes. ... DTU Dataset. The dataset consists of 128 object-scenes...
Dataset Splits No The paper describes training and testing protocols, including ray sampling and steps, but does not explicitly define a separate validation dataset split with specific percentages or counts for hyperparameter tuning.
Hardware Specification Yes Our model is implemented using Py Torch 1.11.0 and all experiments are conducted on Nvidia RTX 3090 GPUs with CUDA 11.4.
Software Dependencies Yes Our model is implemented using Py Torch 1.11.0 and all experiments are conducted on Nvidia RTX 3090 GPUs with CUDA 11.4.
Experiment Setup Yes In our experiments, λ1 is set as 0.1 and λ2 is set as 1. ... In Setting I, we also employ depth maps (Liu et al., 2022) as priors to assist the Hypernet modules to generate adaptive weights. Following (Liu et al., 2022), we randomly sample 2048 rays from Target-Reference pairs, and it trains for a total of 600,000 steps. In Setting II (Wang et al., 2022), we randomly sample 512 rays from Target-Reference pairs, and it trains for a total of 400,000 steps without any priors. In order to enhance training and inference efficiency, we sample K = 64 points along each ray and simplify the volume rendering process in our paradigm.